Bug 295575 – [LibWebRTC, arm64] OPUS codec lacks NEON optimization on Apple Silicon

• Component: WebRTC
• Status: RESOLVED FIXED

[Yury Yarashevich]

Created attachment 475817 [details]
Flame graph of webrtc::AudioEncoderOpusImpl::EncodeImpl(unsigned int, rtc::ArrayView<short const, -4711l>, rtc::BufferT<unsigned char, false>*)

**Description**:

When WebRTC is compiled as part of WebKit (see https://github.com/WebKit/WebKit/blob/main/Source/ThirdParty/libwebrtc/CMakeLists.txt), the OPUS codec is built to operate using only the plain C implementation of all functions, rather than using the optimized versions written with NEON intrinsics.

This causes unnecessary CPU usage and degraded performance during real-time audio compression/decompression.

**Steps to reproduce**:
1. Open https://webrtc.github.io/samples/src/content/peerconnection/pc1/ and establish "call";
2. Capture CPU profile of Safari's processes with Instruments;

**Actual result**:
OPUS encoding uses fallback C implementations of functions like `silk_NSQ_del_dec_c`, `clt_mdct_forward_c` and `silk_LPC_inverse_pred_gain_c`.

**Expected result**:
OPUS encoding uses optimized NEON implementations like `silk_NSQ_del_dec_neon`, `op_pvq_search_c`, `clt_mdct_forward_neon` and `silk_LPC_inverse_pred_gain_neon`

**Environment**:
1. Safari 18.5 on MacOS 15.5 (Apple M2 Pro);
2. Safari Technology Preview Release 222 (WebKit 20622.1.17.2)  on MacOS 15.5 (Apple M2 Pro);
3. Safari 18.5 on iOS 18.5 (iPhone 11).
a

[Alexey Proskuryakov]

Is this the same as bug 169884?

[Yury Yarashevich]

I've only noticed the existing issue after I've created this one, but it seems like original issue may be at least partially solved: https://github.com/WebKit/WebKit/blob/8afd08557be462ffe1b3cf06cd39c806dcf37afc/Source/ThirdParty/libwebrtc/CMakeLists.txt#L2649-L2657, but I don't x86_64 machine to check.

[Sam Sneddon [:gsnedders]]

(In reply to Alexey Proskuryakov from comment #1)
> Is this the same as bug 169884?

Yes, and the resolved bug 170592.

(In reply to Yury Yarashevich from comment #2)
> I've only noticed the existing issue after I've created this one, but it
> seems like original issue may be at least partially solved:
> https://github.com/WebKit/WebKit/blob/
> 8afd08557be462ffe1b3cf06cd39c806dcf37afc/Source/ThirdParty/libwebrtc/
> CMakeLists.txt#L2649-L2657, but I don't x86_64 machine to check.

The Apple ports don't use cmake.

Looking at the code, I wonder if this is because we define OPUS_ARM_PRESUME_AARCH64_NEON_INTR but not OPUS_ARM_PRESUME_NEON_INTR — so we only get A64-only NEON instructions, not the ones shared with A32.

*** This bug has been marked as a duplicate of bug 169884 ***

[Sam Sneddon [:gsnedders]]

(I didn't mean to close this as a duplicate after noticing the above.)

[Yury Yarashevich]

Pull request: https://github.com/WebKit/WebKit/pull/47884

[Radar WebKit Bug Importer]

<rdar://problem/155851882>

[EWS]

Committed 298080@main (bc484ce61b8e): <https://commits.webkit.org/298080@main>

Reviewed commits have been landed. Closing PR #47884 and removing active labels.

[WebKit Commit Bot]

Re-opened since this is blocked by bug 296744

[Yury Yarashevich]

Pull request: https://github.com/WebKit/WebKit/pull/48818

[EWS]

Committed 298886@main (5ddd462f59cd): <https://commits.webkit.org/298886@main>

Reviewed commits have been landed. Closing PR #48818 and removing active labels.